On 2018-09-27 13:01, Geert Uytterhoeven wrote: > Hi Rasmus, > > On Thu, Sep 27, 2018 at 12:55 PM Rasmus Villemoes > <linux@xxxxxxxxxxxxxxxxxx> wrote: >> On 2018-09-27 01:13, Kees Cook wrote: >> >> Just drop devm_kfree_const and teach devm_kfree to ignore >> is_kernel_rodata(). That avoids the 50-100 bytes of overhead for adding >> yet another EXPORT_SYMBOL and makes it easier to port drivers to >> devm_kstrdup_const (and avoids the bugs Kees is worried about). devm >> managed resources are almost never freed explicitly, so that single >> extra comparison in devm_kfree shouldn't matter for performance. > > I guess we can also teach kfree() to ignore is_kernel_rodata(), and > drop kfree_const()? In principle, yes, but the difference is that kfree() is called a lot more frequently, and on normal code paths, whereas devm_kfree is more often (though not always) called on error paths. The goal of _const variants of strdup is to save some memory, so one place to start is to reduce the .text overhead of that feature. And it avoids introducing subtle bugs if some devm_kfree() call is missed during conversion to devm_kstrdup_const(). Rasmus
